You've already forked ComfyUI-WanVideoWrapper
mirror of
https://github.com/kijai/ComfyUI-WanVideoWrapper.git
synced 2026-01-26 23:41:35 +03:00
863829d083
Unsure when this has changed but seems that _scaled_mm now works with same dtype? This fixes the quality degradation of fp8_fast in initial tests.
40 lines
1.7 KiB
Python
40 lines
1.7 KiB
Python
#based on ComfyUI's and MinusZoneAI's fp8_linear optimization
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
def fp8_linear_forward(cls, original_dtype, input):
|
|
weight_dtype = cls.weight.dtype
|
|
if weight_dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
|
|
if len(input.shape) == 3:
|
|
#target_dtype = torch.float8_e5m2 if weight_dtype == torch.float8_e4m3fn else torch.float8_e4m3fn
|
|
inn = input.reshape(-1, input.shape[2]).to(weight_dtype)
|
|
w = cls.weight.t()
|
|
|
|
scale = torch.ones((1), device=input.device, dtype=torch.float32)
|
|
bias = cls.bias.to(original_dtype) if cls.bias is not None else None
|
|
|
|
if bias is not None:
|
|
o = torch._scaled_mm(inn, w, out_dtype=original_dtype, bias=bias, scale_a=scale, scale_b=scale)
|
|
else:
|
|
o = torch._scaled_mm(inn, w, out_dtype=original_dtype, scale_a=scale, scale_b=scale)
|
|
|
|
if isinstance(o, tuple):
|
|
o = o[0]
|
|
|
|
return o.reshape((-1, input.shape[1], cls.weight.shape[0]))
|
|
else:
|
|
return cls.original_forward(input.to(original_dtype))
|
|
else:
|
|
return cls.original_forward(input)
|
|
|
|
def convert_fp8_linear(module, original_dtype, params_to_keep={}):
|
|
setattr(module, "fp8_matmul_enabled", True)
|
|
|
|
for name, module in module.named_modules():
|
|
if not any(keyword in name for keyword in params_to_keep):
|
|
if isinstance(module, nn.Linear):
|
|
original_forward = module.forward
|
|
setattr(module, "original_forward", original_forward)
|
|
setattr(module, "forward", lambda input, m=module: fp8_linear_forward(m, original_dtype, input))
|